Preview Match Score (PMS)

• Preview Match Score (PMS) is a metric that synthesizes quantitative and qualitative data to predict or contextualize competitive outcomes.
• It combines pre-match statistics, live match data, expert commentary, and physiological signals using statistical classifiers and ensemble models.
• PMS is applied across domains from sports broadcasting and betting to clinical assessments and AI evaluations, enhancing real-time decision-making.

A Preview Match Score (PMS) is a quantitative or qualitative metric that predicts, summarizes, or contextualizes likely match outcomes or in-game developments before or during a sporting contest. PMS frameworks have been implemented and studied in esports, traditional sports analytics, physiological assessment, and evaluation of AI models, serving both broadcast and analytical functions. PMS concepts unify diverse modeling strategies—from pre-match statistical ratings and live time-series predictions to integrated objective/subjective scoring—under the goal of providing interpretable, actionable previews of competitive outcomes.

1. Core Concepts and Definitions

The main objective of a Preview Match Score is to synthesize available data—either pre-match or dynamically updating during the contest—into an interpretable measure that forecasts or characterizes the unfolding competitive balance. PMS systems typically operationalize this in one of two forms:

• Probability-based PMS: Using predictive models (e.g., machine learning classifiers or statistical frameworks) to express a team's or player's win likelihood as a percentage or score.
• Composite indices: Merging multiple feature sources (e.g., physiological signals, subjective assessments, or expert analysis) into a single or multi-dimensional score.

In esports, PMS has been equated with "win prediction" using machine learning methods whose outputs (such as predicted probabilities or intuitive game statistics) are suitable for real-time broadcast overlays or descriptive analytics (Hodge et al., 2017). In sports such as football, PMS frameworks blend predicted match statistics, market-derived information, and contextual features from expert commentary to enhance match forecasts (Wheatcroft, 2020, Beal et al., 2020).

2. Methodologies for PMS Construction

Data Sources and Feature Engineering

PMS is fundamentally data-driven, with feature selection tailored to the domain:

• Pre-match features: In MOBA esports, hero selection vectors encode the pre-game draft ((Hodge et al., 2017), e.g., $x_i = +1$ for Radiant, $-1$ for Dire, $0$ otherwise), representing the strategic landscape before play begins.
• In-game or real-time features: Match statistics (kills, damage, net worth) are acquired via time-series slices, enabling dynamic PMS updates as the contest progresses.
• Predicted statistics: Football PMS models utilize Generalised Attacking Performance (GAP) ratings to generate pre-match expectations for shots, corners, or similar metrics (Wheatcroft, 2020).
• Textual and contextual features: NLP techniques extract qualitative insights from expert-written match previews, mapping narrative content into structured vectors (Beal et al., 2020).
• Physiological and psychological features: In clinical domains, PMS can integrate objective neural activity measures (e.g., NIRS-derived Oxy-Hb integrals) with subjective mood scales to evaluate states such as premenstrual syndrome (Aoki et al., 10 May 2024).

Modeling Approaches

Common modeling strategies for PMS include:

• Statistical classifiers: Logistic Regression and Random Forests, trained on engineered features, estimate win probabilities or categorical outcomes ((Hodge et al., 2017); $P(y=1|X) = \frac{1}{1 + e^{-\beta^T X}}$).
• Ordinal logistic regression: Integrates predicted match statistics and betting odds to generate richer PMS distributions for multi-outcome sports (Wheatcroft, 2020).
• Latent variable models: Gaussian Process frameworks model running score differences, generating probabilistic indices of team momentum and match excitement that may inform PMS (Ekstrøm et al., 2020).
• Ensemble and hybrid models: Combining statistical predictions, bookmaker odds, and expert-derived features via ensemble architectures improves forecast accuracy (Beal et al., 2020).

Feature Selection and Optimization

Feature selection methodologies are applied contextually:

Data Type	Feature Selection Approach	Notes
Hero vectors	Wrapper-based (e.g., BestFirst)	Higher accuracy with wrapper
In-game stats	Filter-based (CfsSubsetEval)	Robust for time-series features
Text vectors	Count Vectorizer, aggregation	Allocated by reference to team/context

Hyperparameter tuning and model validation (e.g., regularization in LR, tree count in RF) are essential for maximizing predictive accuracy and robustness.

3. Applications Across Domains

Broadcast and Spectator Engagement

PMS outputs—especially win probabilities or simple statistics such as kill differential ($\text{Kills}_{R-D}$)—are suitable for real-time display, allowing broadcasters to communicate shifting game advantage and anticipated outcomes to audiences (Hodge et al., 2017). Dynamic PMS enables "live" momentum scoring, enhancing interpretability of complex gameplay.

Gambling and Betting Models

PMS frameworks that incorporate predicted match statistics and betting market information facilitate advanced betting strategies:

• Level Stakes Value Betting: Bets placed when predicted PMS probability exceeds that implied by odds.
• Kelly Criterion: Stake size optimized in proportion to the edge between PMS-derived probabilities and odds-implied probabilities (Wheatcroft, 2020).

Empirical studies demonstrate that integrating PMS with odds delivers additional information beyond what is captured by market prices alone, though diminishing returns are possible as market efficiency improves.

Clinical and Biomedical Assessment

In the context of emotional disorders, PMS can be operationalized as a composite score integrating physiological markers (e.g., NIRS Oxy-Hb responses) and standardized questionnaire results (e.g., POMS2), supporting objective tracking of syndromic status or response to interventions (Aoki et al., 10 May 2024). The composite scoring can be formalized as:

$$\text{PMS} = \alpha \cdot I + \beta \cdot (\text{POMS2}_\text{total})$$

where $I$ is the physiological integral and coefficients are fitted through statistical modeling.

AI and Model Evaluation

The PMS concept extends naturally to AI systems and model evaluation. The Performance Rating Equilibrium (PRE) is related in spirit, defining a vector of self-consistent ratings that would perfectly "preview" tournament results given the actual outcomes—a fixed point conceptually adjacent to PMS but formalized via multidimensional equilibrium (Ismail, 21 Oct 2024).

4. Empirical Findings and Validation

PMS systems demonstrate varying accuracy and utility across domains:

• Esports: Mixed-rank models achieved approximately 76.2% win prediction accuracy using in-game time-series data; pure hero-selection models were less predictive for professional matches, highlighting the greater unpredictability of pro-level strategies (Hodge et al., 2017).
• Football: PMS models based on GAP ratings and logistic regression identified profitable betting opportunities and improved forecast accuracy, with particular strength in measures like predicted shots off target (Wheatcroft, 2020).
• Combined statistical and human insight: Ensemble models including text-derived features achieved a 6.9% accuracy boost over classical statistical methods in football match outcome prediction (Beal et al., 2020).
• Physiological and mood-based PMS: Statistically significant differences in PMS (composite of NIRS and POMS2) distinguished PMS and PMDD from control groups, supporting validity as an objective assessment (Aoki et al., 10 May 2024).

5. Limitations and Challenges

Several recurring limitations affect PMS systems:

• Data scarcity and heterogeneity: For professional esports, limited occurrences necessitate supplementing with high-level non-pro data, which introduces performance degradation due to style differences (Hodge et al., 2017).
• Feature drift and data obsolescence: Rapid changes in meta-games or sports tactics can render historical data less predictive, requiring frequent recalibration of PMS models.
• Model robustness and transferability: Optimal PMS configurations may not generalize across domains or player populations—feature selection and model tuning must be context-specific.
• Complexity and interpretability: While advanced models (e.g., latent GPs for momentum) offer fine-grained dynamic PMS, computational cost and decreased transparency can hinder real-time deployment and user understanding (Ekstrøm et al., 2020).
• Evaluation consistency: Traditional n-gram-based metrics perform poorly for long-form, context-rich evaluations—bespoke frameworks such as Extract, Match, and Score (EMS) offer improved alignment with practitioner priorities (Hu et al., 20 Mar 2025).

6. Methodological Advancements and Future Directions

Recent research signals several promising trajectories for PMS:

• Integration of heterogeneous data sources: Blending subjective expert narratives, statistical performance indicators, physiological measures, and even market sentiment to create richer, more holistic PMS frameworks (Beal et al., 2020, Wheatcroft, 2020, Aoki et al., 10 May 2024).
• Dynamic and context-aware metrics: Latent state modeling and real-time Bayesian updating enable PMS systems to track not just outcome likelihood, but also match "excitement" and turning points (Ekstrøm et al., 2020).
• Self-consistency and equilibrium-based ratings: Performance Rating Equilibrium introduces true fixed-point PMS, applicable to both sports and AI model assessments, supporting fairness and internal consistency in retrospective evaluation (Ismail, 21 Oct 2024).
• Granular and actionable feedback: EMS frameworks systematically evaluate factual completeness and precision for long-form outputs, providing fine-grained PMS-style diagnostics in domains such as financial analysis (Hu et al., 20 Mar 2025).

A plausible implication is that future PMS systems will be increasingly multimodal, context-adaptive, and interpretable, serving diverse stakeholders from spectators to analysts to clinical practitioners. Careful validation, regular recalibration, and transparency of modeling assumptions remain vital for their sustained impact and reliability.